Nace AI Launches Drex for Search, Achieving 92.05% Accuracy on BrowseComp-Plus

27 October 2026 3 min read هدهد تك | Hudhud Tech

Nace AI has released Drex for Search, integrating its Drex decision models into document search. This new retrieval layer assists users and agents in locating evidence efficiently. Coding agents such as Claude Code, Codex, Cursor, and OpenCode can access it via CLI, SDKs, MCP, or REST. In Nace’s evaluation, Drex for Search scored 92.05% on BrowseComp-Plus using GPT-5, with an average of 12.6 search calls per question.

Key Features

Size: A hosted search system designed to search over 1 million files, compatible with answering models like GPT-5, Gemini 3.5 Flash, or Qwen 27B.
Deployment: Operates on Nace’s managed cloud, a dedicated cloud, personal AWS or Azure accounts, or on-premises setups.
Performance: Ranked #1 or #2 on three out of four public leaderboards, according to Nace.
Best Performance: Achieved 77.5% R@1 on VisDocAgentBench, outperforming Claude Opus 5 by 10 points.

Overview of Drex for Search

Nace AI positions search as a decision-making process. Drex for Search applies its models to evaluate options, assigning calibrated probabilities in a single forward pass, making it cost-effective for each agent step.

The release focuses on classic Retrieval-Augmented Generation (RAG), summarized by Nace as chunk, embed, top-k, hope.

Pipeline Limitations:

When the answer is located in a table on page 23.
If the answer requires two hops across files.
When the answer is embedded within a scanned PDF.

Functionality Within an Agent

Nace outlines three operational steps:

Classify: Drex assesses whether the question requires a quick lookup or an agent. The agent route involves reading documents and querying tables.
Squeeze: Drex navigates through the workspace level by level: folder, file, section, and page, narrowing down to the relevant table and column.
Answer: Only the retained pages and tables are sent to the LLM, with the answer referencing the specific file and page.

Drex maintains a fixed depth, meaning that processing a million files requires the same number of steps as processing a thousand.

Impact on Coding Agents

Nace AI evaluated nine agent and model combinations using OfficeQA Pro, a benchmark consisting of 133 questions based on U.S. Treasury Bulletins. The results showed a 6% increase in correct answers, a 62% reduction in steps, a 56% decrease in tokens used, and a speed improvement of 2.3 times.

Workspace Components

Files: Supports various formats including PDFs, scans, xlsx, csv, Parquet, mp3, mp4, docx, pptx, md, json, and txt.
Citations: Each answer specifies the file and the page, row, or minute from which it was derived.
Workspaces: Each workspace is isolated with its own index and knowledge graph, shareable with invited members.
Storage: Users can upload files or folders, or store workspaces in their own S3 bucket or Azure Blob container.
Zero Retention: Available with personal storage solutions.
Domains: Includes Tax, M&A, SEC 10-K, Contracts, or custom domains.

Performance on BrowseComp-Plus

BrowseComp-Plus challenges a deep-research agent with 830 multi-hop questions across approximately 100,000 fixed web documents. With GPT-5, Nace AI reports:

92.05% accuracy, ranking #2 out of 86 entries on the leaderboard.
12.61 search calls per question, significantly lower than the leading entry, which required 209.27 calls.
87.32% accuracy in a single pass with 8.93 search calls, the best performance among entries under 10 calls.
+20.4 points over the Qwen3-Embed-8B baseline, using the same GPT-5 model.